🐓 Bawk Sampler

This is the node the whole pack is built around. Bawk Sampler does what a stock ComfyUI setup needs five nodes to do - empty latent, ModelSamplingFlux, KSampler, VAEDecode, plus the wiring between them - and collapses it into one. It's a FLUX-tuned, all-in-one txt2img/img2img sampler that hands you finished images on the other end, which is exactly the "no more node spaghetti" pitch the pack's author made when he posted the pack to r/comfyui.

How it works

Under the hood it's not magic, just assembly. It generates an empty latent (16-channel, FLUX's flavor) or encodes your input_image if you provide one, patches the model with ModelSamplingFlux using max_shift/base_shift (that's the flow-matching sigma-shift knob - ComfyUI's way of distributing sampling effort between composition and detail), injects guidance into the conditioning, runs the sampler/scheduler you picked, and decodes with your VAE. The latent output is still there if you want a second pass or a hires-fix chain.

The settings that matter

• resolution - 14 presets. 1024×1024 is the FLUX sweet spot and the default. 512×512 "Quick Test" for drafts, 1408×1408 "FLUX Max Quality" for the official 2MP ceiling (that one wants serious VRAM), plus social sizes (Instagram Square, Instagram Portrait, Twitter Post). Need something else? Flip use_custom_resolution and set custom_width/custom_height.
• batch_size - default 4. That's four images per run; on a 12GB card consider dropping to 1–2 while you iterate.
• sampler/scheduler - euler + beta defaults are the right FLUX call. The community rule still holds: never Karras for FLUX - it distorts the flow-matching trajectory and just makes things worse.
• steps - 30 default, 20 for speed, 40 for quality.
• guidance - 3.5 default (balanced), 2.0 creative, 7.0 structured. That's FLUX's guidance-in-conditioning scheme, not old-school CFG.
• max_shift/base_shift - 1.15/0.5 defaults are the community-tested FLUX values. Higher max_shift = more coherent, lower = more varied.
• denoise - 1.0 for txt2img; 0.6–0.9 for img2img via input_image.
• use_custom_resolution + custom_width/custom_height - when presets aren't enough. Dimensions must be multiples of 64, and the node rounds for you.

Outputs and wiring

images (already VAE-decoded - wire straight into Bawk Image Saver) and latent (for img2img or upscale passes). The pack's quick-start chain: Bawk Model Loader → Bawk Wildcard Encoder → Bawk Sampler → Bawk Image Saver.

Install & troubleshooting

Same pack install - ComfyUI Manager → search "Bawk Nodes" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/juddisjudd/ComfyUI-BawkNodes.git

The pack's requirements.txt adds nothing beyond stock ComfyUI (torch, numpy, safetensors, requests), and the sampler's logic uses ComfyUI's own comfy_extras nodes, so there's no version drift to worry about.

Where people trip up:

• It's FLUX-tuned by default. Point it at Z-Image and the README's rule applies: steps 8, guidance 0.0. Leave it at 30/3.5 and a distilled model like Z-Image Turbo will over-correct and burn the image - guidance is baked into those weights.
• batch_size 4 surprises people. If you get an OOM on the first run, that's usually it.
• The default resolution presets are chosen for FLUX's ~2MP ceiling. Above that, generate at 1408 and upscale rather than cranking the latent size.